AILang/docs/journals/2026-05-12-audit-23.md

# audit-23 — Milestone 23 close

**Date:** 2026-05-12
**Started from:** 3fed372
**Status:** DONE — milestone 23 closed (1 ratify + 1 nit-fix)
**Skill:** audit (architect + bencher)

## Summary

Milestone-close audit for Milestone 23 (Eq/Ord prelude). Architect
clean on substance — all 8 acceptance criteria of
`docs/specs/2026-05-11-23-eq-ord-prelude.md` met, no out-of-scope
creep, DESIGN.md two-arm-fixpoint description (line 1513-1556)
matches `mono.rs::collect_mono_targets` observable behaviour, roadmap
split internally consistent. Bencher found and ratified one
structural compile-time shift; the iter-23.5 journal's "one sub-ms
noise-band metric" framing was wrong on both counts (4 metrics, not
1; consistent +18-20% cohort shift, not noise).

## Architect — drift items

- **[important]** `docs/DESIGN.md:1692-1693` — sentence collision:
  "Milestone 22 ships **no built-in Prelude classes**. The original
  An earlier draft committed to …". Pre-existing since commit
  `613d4d8` (2026-05-10, design-md-consolidation 1.4 fixup) which
  edited `22a draft` → `An earlier draft` and left the dangling
  fragment `The original` on the previous line. NOT introduced in
  iter 23.5; architect's attribution corrected. Boss-fixed inline:
  the dangling fragment removed.
- **[minor]** `crates/ailang-check/src/lib.rs:884-894` — workspace-
  sibling linearity construction is O(N²) over module count. Bounded
  for today's 2-module workspace; named here so a future Nx-module
  workspace doesn't re-discover it. Bencher confirmed below that the
  iter-23.5 H1 cost is negligible at N=2. No action.
- **[minor]** `crates/ailang-check/src/mono.rs:703-712` —
  `contains_rigid_var` discriminates rigid vs. metavar by
  `Type::Var.name.starts_with("$m")`. The journal already named this
  as a future-fragility; a structural discriminator (separate
  `Type::Meta` variant) would be the principled fix. Carry-on; not
  blocking.
- **[nit]** `crates/ailang-check/src/lib.rs:709-726` — Float-NoInstance
  addendum mutates `d.message` directly. The codebase has no
  `with_hint` / `extend_message` convention yet, so this isn't
  pre-existing drift, but the next class-specific addendum (e.g.
  partial-Float `compare`) would compound `to_diagnostic` into a
  conditional chain. Carry-on; refactor when a second addendum
  arrives.

## Bencher — compile_check regression diagnosis

Bench-rerun at HEAD `3fed372` showed 4/12 check_ms metrics above
the 25% tolerance, with all 12 metrics in the +7 to +45% band — a
coherent cohort shift, not the journal-claimed "one sub-ms noise
metric".

Two hypotheses tested via 5-run medians over 60 samples each across
four states (23.4 HEAD, 23.5 HEAD, 23.4 checker + 23.5 prelude,
23.5 checker + 23.4 prelude):

- **H1** (linearity workspace-visibility + `contains_rigid_var`):
  refuted. 23.5-checker + 23.4-prelude returns to baseline noise
  (median +1.84% vs. baseline noise floor).
- **H2** (5 new polymorphic prelude `Def::Fn` bodies cost real
  typecheck work): confirmed. 23.4-checker + 23.5-prelude reproduces
  the full HEAD regression (median +18.62% vs. HEAD's +18.74%; gap
  bounds the H1 interaction at <1%).

Cost source: each workspace check now typechecks 5 additional
polymorphic `Def::Fn`s (each carrying `forall` + 1 class constraint
+ a `match` or small `not (eq …)` body) plus their constraint
resolution. ~0.2-0.3ms median additional work per `ail check`
subprocess; the ~5-10ms subprocess-spawn floor inflates the
relative-percent figure.

**Verdict:** Ratify. The regression is the user-visible feature iter
23.5 shipped — five prelude helpers — accurately surfacing in the
bench as "typecheck does more work because the prelude has more
work to do".

## Resolution

- **Ratify path applied:** `python3 bench/compile_check.py
  --update-baseline` re-measured (1 run, default 5 samples) and
  wrote `bench/baseline_compile.json` with the new check_ms numbers.
  Post-update rerun: 0/24 regressed, 0/24 improved beyond tolerance.
- **Important-drift fix applied:** the dangling `The original`
  fragment removed from `docs/DESIGN.md:1692`. One-line tidy edit;
  no behavioural change.
- **Minor / nit drift items:** carry-on. Linearity O(N²) is
  observed-not-blocking; mono name-prefix discriminator is journal-
  flagged; NoInstance addendum mutation pattern revisit when a
  second class-specific addendum arrives.

## Concerns

- The iter-23.5 journal's bench framing was incorrect. The
  implementer asserted "one sub-ms metric, two others stayed inside
  on rerun" — fresh-rerun showed 4 above tolerance with consistent
  cohort shift. Implementer-phase bench rerun count was likely too
  small to see the cohort; orchestrator-audit pattern of running
  bench 5× back-to-back at milestone close is the correct
  granularity, and the 1-sample bench-check inside `implement` is
  not.
- Implication for future iters: when `compile_check.py` exits 1
  during implement, the implementer should NOT classify it as
  noise-band without doing the same multi-run, multi-state isolation
  the bencher did here. Defer to audit if the framing is uncertain.
  This is a soft convention — encoding it as a hard skill rule would
  slow implement when the regression genuinely is noise; the
  weaker form is "if bench exits 1, the journal Concerns must
  describe what additional evidence was gathered, not just give a
  hypothesis".

## Known debt

- DESIGN.md sentence-collision had been on main for two days
  unnoticed. No drift-review ran between `613d4d8` (2026-05-10)
  and this audit (2026-05-12); the implement-skill diff scope
  (per-iter, narrow) does not catch unrelated pre-existing
  fragments. Acceptable: drift-review at milestone close is exactly
  the catch mechanism, and it caught it.

## Files touched

- `bench/baseline_compile.json` — re-measured baseline; numbers
  reflect new prelude cost.
- `docs/DESIGN.md:1692` — dangling fragment removed.
- `docs/journals/2026-05-12-audit-23.md` (this file).
- `docs/journals/INDEX.md` — append (Boss-side).

## Stats

No stats file — audit is a Boss-direct skill, not an
implement-orchestrator dispatch.